Automatic CT Angiography Lesion Segmentation Compared to CT Perfusion in Ischemic Stroke Detection: a Feasibility...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• CT angiography (CTA) is crucial for identifying arterial occlusions in stroke patients.
• CTA images hold potential for assessing the extent of ischemic damage.
• Accurate and rapid assessment of ischemic lesions is vital for effective stroke treatment.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based algorithm for detecting and segmenting acute ischemic lesions from CTA.
• To compare the CNN algorithm's performance against manual segmentations and established CT perfusion software (RAPID).
• To assess the clinical applicability of the CNN algorithm by analyzing false positives and negatives.

Main Methods:

• A 42-layer deep CNN was trained on 50 CTA volumes with manually delineated ischemic lesions.
• The algorithm's ability to differentiate stroke from non-stroke was evaluated based on predicted lesion size.
• Visual review of false positives and negatives was conducted to assess clinical utility.

Main Results:

• The CNN model achieved a voxel-wise sensitivity of 0.54 and a Sørensen-Dice coefficient of 0.61 compared to manual segmentations.
• An accuracy of 0.88 was achieved for stroke/non-stroke differentiation based on lesion size.
• The CNN model showed a moderate positive correlation (Pearson's r = 0.76) with RAPID-reported Tmax > 10s volumes.

Conclusions:

• A CNN-based algorithm shows feasibility for detecting anterior circulation ischemic strokes from CTA.
• Clinical applicability requires integration with physiological knowledge to mitigate false positives.
• The algorithm demonstrates potential as an adjunct tool in stroke imaging analysis.